Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 2 Apr 2020 12:08:20 -0300
From:   Marcelo Schmitt <marcelo.schmitt1@gmail.com>
To:     linux-kernel@vger.kernel.org
Cc:     tglx@linutronix.de, lars@metafoo.de
Subject: [RFC] genirq: prevent allocated_irqs from being smaller than NR_IRQS
Message-ID: <20200402150820.GB5886@smtp.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hi,

I was trying to understand IRQ initialization when suddenly got
intrigued about the declaration of the "allocated_irqs" bitmap at
kernel/irq/irqdesc.c. The size of allocated_irqs is defined by
IRQ_BITMAP_BITS, which in turn is passed to BITS_TO_LONGS to calculate
the actual number of IRQs the system may have. If I got it right, there
should be one entry at allocated_irqs for each possible IRQ line. At
kernel/irq/internals.h, IRQ_BITMAP_BITS is defined to be NR_IRQS (or
NR_IRQS plus a high constant in the case of sparse IRQs), which most
architectures seem to define as being the actual number of IRQs a board
has.

#ifdef CONFIG_SPARSE_IRQ
# define IRQ_BITMAP_BITS (NR_IRQS + 8196)
#else
# define IRQ_BITMAP_BITS NR_IRQS
#endif

The thing I'm troubled about is that BITS_TO_LONGS divides
IRQ_BITMAP_BITS by sizeof(long) * 8, which makes it possible for the
size of allocated_irqs to be smaller than NR_IRQS.

For instance, if !CONFIG_SPARSE_IRQ, sizeof(long) == 8, and NR_IRQS is
defined as 16, then IRQ_BITMAP_BITS would be equal to 
(16 + 64 - 1)/64 = 1. Even if CONFIG_SPARSE_IRQ is defined, a device
with a large number of IRQ lines would end up with a small bitmap for
allocated_irqs.

I thought NR_IRQS would be multiplied by the number of bits it uses.
Something like:

#ifdef CONFIG_SPARSE_IRQ
# define IRQ_BITMAP_BITS (NR_IRQS * BITS_PER_TYPE(long) + 8196)
#else
# define IRQ_BITMAP_BITS (NR_IRQS * BITS_PER_TYPE(long))
#endif

Anyhow, IRQ_BITMAP_BITS is also used to limit the maximum number of IRQs
at irqdesc.c. If my understanding of nr_irqs is correct, it would make
sense to change some sanity checks at early_irq_init() too.

Does anyone mind giving me some advice on how allocated_irqs is
initialized with a suitable size to support the number of interrupt
lines a board may have?

By the way, is there any mailing list for IRQ related discussions?
I couldn't find one at vger.kernel.org.


Thanks,

Marcelo